Shadows of Reality

Large language models feel intelligent because they are fluent, but their fluency hides a basic constraint: they do not engage with the world itself. They engage with descriptions of the world. Language is already an abstraction, a compressed record of reality filtered through human perception. When an LLM learns, it is not discovering gravity or motion. It is learning how humans describe those things.

This is why their limitations appear in subtle ways. They can model causality within the representations they are given, but those representations remain narrower than the full dimensionality of the physical world. Their knowledge comes from patterns in abstraction rather than continuous interaction with reality itself. They predict what should come next in language, not necessarily what must happen next in the world.

The obvious response is to give machines access to richer forms of experience. This is the promise of world models. Systems that see, act, and learn through feedback begin to internalize structure rather than merely description. Instead of modeling language about reality, they begin modeling parts of reality directly.

This shift matters because modern machine learning is already built on a mathematical language deeply tied to how we model the physical world itself. As MIT professor Gilbert Strang often emphasizes, linear algebra is not just a branch of mathematics. It is one of the clearest frameworks we have for representing transformations, constraints, and relationships across complex systems. Physics, engineering, graphics, and machine learning all reduce reality into vectors, matrices, and operators that describe how states evolve and interact.

Neural networks themselves are ultimately enormous systems of linear transformations layered together. In that sense, machine learning already operates through the same representational machinery we use to model the structure of reality. The limitation is not that these systems lack mathematics or structure. It is that their access to reality remains partial. A language model learns transformations over text. A world model begins learning transformations grounded in sensory interaction and physical constraint.

It is tempting to see this as an escape from abstraction into reality. But Plato’s cave still applies. A language model operates on shadows cast by human discourse, learning their structure with extraordinary precision. A world model steps closer to the source, grounding itself in sensory input and feedback. But it still constructs an internal representation of the world rather than accessing the world directly.

Humans do the same. Our perception feels immediate, yet it is also mediated through models built from sensory input, memory, and cognition. We do not navigate reality directly. We navigate an internal reconstruction of it.

The difference, then, is not between being inside the cave and escaping it. It is between inhabiting smaller and larger caves. Language models operate inside a narrow cave built from text and representation. World models expand that cave through interaction and constraint. Human cognition expands it further still.

And increasingly, those caves are beginning to converge. The internal worlds machines build to navigate reality are becoming less distinguishable from the ones humans construct for themselves. If linear algebra is the language through which both physics and machine learning represent the world, then AI is not inventing a separate reality so much as building increasingly rich caves from the same underlying structure. Not because machines have escaped representation, but because they are learning to inhabit it in ways that increasingly resemble us.